Abstract: A spectral-division unsupervised shot-clustering algorithm (SDUSC) was proposed. First, the keyframes were picked out to represent the shots, and the color features of keyframes were extracted to describe the video shots. In order to characterize shots optimally, a spherical gaussian model (SGM) was constructed for every shot category. Then the spectral division (SD) method was employed to divide a category into two categories, and the method was iteratively used to further divide the results of previous divisions. At the end of each iterative shot-division, bayesian information criterion (BIC) was utilized to automatically judge whether to stop further division. During the processes of shot-divisions, one category might be dissevered by mistake. In order to correct these mistakes, similar categories would be merged by calculating the similarities of all the possible category pairs. This approach was applied to three kinds of sports videos, and the experimental results shows that the proposed approach is reliable and effective.